Claude Fable 5
by Anthropic
Anthropic’s most capable widely-released model. Fable 5 excels at long-horizon agentic reasoning, autonomous planning, complex problem-solving, and multidisciplinary work. Includes safety refusal classifiers and adaptive thinking. Released June 9, 2026.
Overview
Claude Fable 5 is Anthropic’s flagship entry in the Claude 5 generation (released June 2026). It represents a breakthrough in autonomous reasoning and long-horizon task execution, with built-in safety mechanisms and adaptive thinking that allow it to adjust computational effort based on task complexity. Fable 5 is the most capable model widely available to all users.
Performance & Benchmarks
- Terminal-Bench 2.1: 88.0%
- Agents’ Last Exam: Competitive with Claude Opus 5 and superior to GPT-5.5
- Long-horizon agentic reasoning: Strongest in Claude lineup
- Coding & software engineering: Near-frontier performance
- Novel problem-solving: Excellent across domains
Pricing
- Input tokens: $10 per million
- Output tokens: $50 per million
- Availability: Full public release (June 2026); briefly restricted under US export controls (~June 16-July 1, 2026), then restored
Core Capabilities
Adaptive Thinking (Always-On)
- Automatic reasoning effort adjustment based on task complexity
- Transparent reasoning process with no performance penalty
- Configurable thinking depth (standard, medium, high)
- Enables complex multi-step problem-solving
Long-Horizon Agentic Reasoning
- Autonomous task decomposition and planning
- Multi-step reasoning across 30+ hour task chains
- Robust error recovery and replanning
- Superior long-context coherence
Extended Vision & Understanding
- Advanced image analysis and diagram interpretation
- Technical document understanding
- Chart and graph interpretation
- Scientific imagery analysis
Computer Use & Tool Orchestration
- Visual interface understanding and interaction
- Desktop application control
- Web automation and scraping
- Terminal command execution
- Parallel and sequential tool coordination
Multidisciplinary Reasoning
- Cross-domain knowledge integration
- Novel hypothesis generation
- Complex scientific reasoning
- Strategic planning and analysis
Knowledge Work
- Research synthesis and literature review
- Complex document analysis
- Policy and regulation interpretation
- Business strategy and planning
Long Context
- Context window: 1,000,000 tokens (1M)
- Max output tokens: 128,000
- Process entire codebases, research archives, or organizational documents in single interaction
- Efficient handling of complex multi-document reasoning
Use Cases
Software Engineering & Development
- Large-scale architecture redesign
- Multi-repository refactoring
- Complex feature implementation
- Autonomous agent development
- Bug investigation and fixing across large codebases
Research & Academia
- Literature review synthesis across hundreds of papers
- Novel hypothesis generation
- Experimental design assistance
- Data analysis and interpretation
- Writing assistance for academic work
Complex Problem-Solving
- Novel technical challenges
- Multi-domain problem decomposition
- Strategic planning and decision-making
- Risk analysis and scenario planning
Business Intelligence & Strategy
- Competitive analysis across multiple sources
- Market research synthesis
- Strategic recommendations
- Policy impact assessment
Autonomous Systems & Agents
- Multi-step task execution without human intervention
- Robust error handling and replanning
- Complex workflow automation
- Long-running background tasks
Knowledge Management
- Organizational knowledge synthesis
- Internal documentation review
- Compliance and policy analysis
- Institutional memory augmentation
Behavioral Characteristics
- Nuanced, sophisticated reasoning
- Excellent at acknowledging uncertainty
- Strong on complex instruction-following
- Direct, professional communication style
- Superior performance on underspecified problems
- Better at novel problem types unseen in training
Safety & Alignment
- Safety refusal classifiers: Distinguishes harmful requests from beneficial edge cases
- Lower misalignment rates: Substantially improved vs. Claude 4 series
- Constitutional AI v2: Updated 80-page constitution (January 2026)
- Robustness: Excellent prompt-injection resistance
- Hallucination control: Reliable on knowledge-intensive tasks
- Sycophancy reduction: Less susceptible to user agreement bias
Adaptive Thinking Architecture
Fable 5 implements always-on adaptive thinking:
- No token-budget constraint: thinks as much as needed per token spent
- Transparent to the user if requested
- Automatically adjusts depth for complex vs. simple tasks
- Enables “frontier performance on command”
Comparison to Claude Opus 5 & Mythos 5
| Feature | Fable 5 | Opus 5 | Mythos 5 |
|---|---|---|---|
| Release | June 9, 2026 | July 24, 2026 | June 9, 2026 |
| Pricing (In/Out) | 50 | 50 | 50 |
| Safety Refusals | Yes | Yes | No (limited-release) |
| Context Window | 1M tokens | 1M tokens | 1M tokens |
| Max Output | 128K | 128K | 128K |
| Adaptive Thinking | Yes | Yes | Yes |
| Primary Use | Widely available, all users | Balanced performance/cost | Limited access (gov’t contracts) |
| Performance | Frontier | Near-frontier | Frontier |
Platform Availability
- Claude Web App (claude.ai)
- Claude API
- AWS Bedrock
- Google Vertex AI
- Enterprise deployments
- (Briefly offline June 16-July 1, 2026 under US export controls; restored July 1)
Training & Knowledge
- Training cutoff: April 2026
- Multilingual capabilities
- Strong awareness of recent events (through April 2026)
- Production-ready for all environments
When to Use Fable 5
Use Fable 5 when you need:
- Most capable widely-available Anthropic model
- Long-horizon autonomous reasoning
- Novel problem-solving beyond the distribution
- Full access to safety mechanisms
- Robust performance across domains
- Complex multi-step task execution
Consider Opus 5 instead if:
- You need near-Fable performance at lower cost and complexity
- Tasks are more standard/predictable
- Context size needs are less extreme
Consider Mythos 5 if:
- You have special access (limited-release via Project Glasswing)
- You explicitly need model without safety refusal classifiers